0

I have to train a model that can approximate the function used to map the inputs (3) onto a single output (1) with sigmoid activation in the hidden layer and tanh in the output layer.

The data is 8 rows of the input-output pair ((X,Y,Z), SUM) where X,Y and Z are the input and SUM is the output.

The values of X,Y and Z are in different random ranges. Now, I am stuck with the problem of deciding between Normalization and/or Standardization. I have gone through some resources but I found the answers in reference to Clustering and Image Classification.

What should I choose? I mean, if normalize or standardize, should it be done for the entire global data (X,Y,Z SUM) or each done differently. Also, if I standardize, then I'll have to de-standardize at the end. Isn't this abnormal?

Pe Dro
  • 2,651
  • 3
  • 24
  • 44
  • Take a look at [this post](https://stats.stackexchange.com/questions/10289/whats-the-difference-between-normalization-and-standardization). You'll probably want to go for standardization. Moreover, what do you mean by "... de-standardize at the end. Isn't this abnormal?" – ruancomelli Jan 20 '20 at 12:36
  • @rugortal Once I have the model trained,and I ask it to make the prediction for the input (0,10,5), the input will first be standardized and then the corresponding result will be given. But as you know, the result should be 0+10+5 = 15. But this is not directly obtained by feeding in the standardized values. – Pe Dro Jan 20 '20 at 12:49
  • OK, now I understand. The point here is that, in general, we *do not* normalize (nor standardize) the targets. When you standardize your data, your machine will learn a function *f* that maps *standardized* inputs *(X, Y, Z)* onto the expected target *SUM*. Therefore, new *standardized* inputs should also result in the expected target. – ruancomelli Jan 20 '20 at 13:34
  • [This post](https://stats.stackexchange.com/questions/111467/is-it-necessary-to-scale-the-target-value-in-addition-to-scaling-features-for-re) is really nice for understanding why we normalize the feature vector, but not the targets. – ruancomelli Jan 20 '20 at 13:36

1 Answers1

0

Check the below url as suggested by rugortal. Hope it would answer your question. Let me know if you need any more help

https://stats.stackexchange.com/questions/111467/is-it-necessary-to-scale-the-target-value-in-addition-to-scaling-features-for-re

https://stats.stackexchange.com/questions/10289/whats-the-difference-between-normalization-and-standardization

Running Rabbit
  • 2,634
  • 15
  • 48
  • 69